Understanding Post-Baccalaureate Cultural Gaps: Building Equitable Ecosystems for AI Research and What We Can Learn from US Federal TRIO Programs

David S. Hippocampus
Department of Computer Science
Cranberry-Lemon University
Pittsburgh, PA 15213
[email protected]
Use footnote for providing further information about author (webpage, alternative address)—not for acknowledging funding agencies.

Abstract

This paper reviews (1) unequal representation in STEM and AI, (2) socioeconomic influences on retention of scientists and researchers, and (3) existing educational opportunity programs for disadvantaged backgrounds, with emphasis on Post-Baccalaureate support. We conclude that cultural and skill gaps create a perceived permanent exclusion from participating in leading scientific communities and, while education programs increase representation for early-career scientists, these trends suggest that additional post-educational support and standardization would improve the retention, belonging, and advancement ecosystem for researcher collaboration in AI.

1 Introduction

Jonathan’s first day at Wheaton, he looked up his course syllabi and panicked. He couldn’t afford the books. He also did not tell anyone he couldn’t afford the books, he just never got them. (joffe-walt_2015)

1.1 Survey of Inequality from early education to early career scientists

The inequalities that inhibit research collaboration range from a multifaceted history of research from historically unequal resource distribution (seth2009putting; roy2018science) to the inequitable access to technology education today (muro2018black; bell2019becomes; vachovsky2016toward). Additionally, studies have shown that skill gaps do not fully account for the lack of representation in AI and STEM, as evidenced by the attrition of people from underrepresented and disadvantaged communities after working in research and engineering fields (palmer2011qualitative). Researchers have found that before any skill gap is a culture gap that is challenging to short-circuit. Below are results from a cultural identity study of incoming freshmen at a private university, where students attended an hour-long student discussion about adjusting to college. The study reviewed the first year performance of first-generation students in correlation to the type of discussion they attended at the beginning of the year chang2018subtle; stephens2014closing.

One group attended a session in which panelists talked about their social class background, and how it affected their transition to college. This was called "difference education." Another group attended a session in which social class backgrounds were not highlighted. Among the students who were in the standard session that didn’t highlight social class, first-generation students had significantly worse GPAs. But among those who were in the difference education sessions, first-generation students had pretty much the same GPA as continuing generation students.

1.2 Culture Gaps and the Machine Learning Community

1.3 How Federal TRIO Programs Decode the cultural barriers in science

Much of the work in these areas are focused on early education throughcollege, with less work focused on graduated education, e.g one that leads to a successfully connected researcher in a field like Artificial Intelligence. Some of the longest standing U.S. federally funded educational opportunity programs are classified as TRIO [https://www2.ed.gov/about/offices/list/ope/trio/index.html], which only programmatically supports undergraduate college students towards pursuing Masters and PhDs.

Due to structural inequity, White students, those whose parents attained a university degree, and those from upper/middle class households are more likely to attain a doctoral degree. One federal program, the Ronald E. McNair Post-baccalaureate Achievement Program, provides undergraduates with academic and financial support to help marginalized students enroll and succeed in graduate school. However, little research has examined how this program has helped students attain the ultimate goal of a PhD.

In addition to research funding, programs like McNair provide both skill and cultural gap training. Essential skills for entering post-baccalaureate studies may include being able to perform well in standardized testing like the GREs, understanding the foundations of research, and public speaking. However, McNair programs also provided opportunities to overcome cultural barriers through programs that helped students communicate and understand their research advisors, engage in dinner-time debates and discussions around current events, and go on graduate school visits, after helping these scholars purchase their first business suit [Cite]. Understandably, these culture gaps persist beyond PhD programs, and as researcher careers advance, so does the elitism. While a skillswise proficiency of a field should advance, cultural elitism proves to be a nontrivial and less-acknowledged hindrance in broadening research collaboration.

1.4 What we can learn about Cultural Gaps from the 5 paragraph essay

A better researched example of cultural gaps can be found through studies on learning english as a second language [CITE]. Educators are often torn by the cost-benefit of teaching to a standard 5-paragraph position structure [CITE]. On one hand, it can make writing formulaic and dry [CITE]; however, such formulas, whether explicitly outlined or internalized from speaking English as a first language give traction to those learning the English language later in life [CITE]. While “teaching to the test” may be frowned upon and seen as just going through the motions, having the frameworks and standards not only provides a proxy for evaluating the equitability of an ecosystem, but also help individuals identify cultural gaps on all sides of the divides.

Regarding, standards, a lot of work in tech is being poured into training students to have the skillsets to pass technical interviews [Google Paper]; however, in tandem, research has also shown that “economic connectedness” generates the motivation that automatically overcomes skill gaps [NPR]. We’ve observed that while skills enable early-career scientists and engineers into STEM research, overcoming the culture gap is required for retention, belonging, and an equitably collaborative ecosystem. Furthermore, identifying standards (or structures) for cultural understanding brings forth accessbility to what was only implicitly understood.

2 Discussion

2.1 Better Metrics

The problems outlined in the previous section have complexities and nuances that can be challenging to solve for. To better understand the problems we are solving, we pose the following structural questions:

What does it mean to be a successful ML Researcher? What are the practices required to be a successful scientist in ML? What types of collaborators exist that benefit ML research? What are the minimum requirements for types of research collaborations in Machine Learning?

As we saw with the 5-paragraph essay, rules and laws are approximations of how we want humans to cooperate. Some people are able to internalize these frameworks from a young age and benefit from that their entire career [CITE,2,3]. Less-affluent backgrounds (more so) benefit from a continued outline of standards at every stage of career, making explicit what were implicit barriers of entry to collaboration.

2.2 Support during PhD Program and beyond

While it is helpful to make explicit the collaboration barriers, what does that support look like? We have some examples from more generic research programs. Post Baccalaureate Examples: Post-Bacc Programs CASBS Faculty resources https://www.facultydiversity.org/ Women in ML (organization (https://wimlworkshop.org/) and Twitter account: @WiMLworkshop) GEM fellowship program: https://www.gemfellowship.org/gem-fellowship-program/ Cohere for AI Scholars program: https://txt.cohere.ai/introducing-the-cohere-for-ai-scholars-program-your-research-journey-starts-here/ Twitter: BlackInAI Google Research: https://research.google/outreach/

2.3 Addressing the Bi-directional Culture Gap

The burden of responsibility, however, is not solely for the underrepresented to bridge. Support must also exist for the gaps of the in-crowd and may potentially be more effective in some cases for those gaps to be filled first.

From what we observe in the literature, many programs work towards inspiring talent [hour of code, AI4All, Upward Bound, Precollegiate Pathways], then discovering [Google BOLD internship] and cultivating talent [McNair]; there is less work on retaining talent from early to later career [Mcnair PhDs, Women of Color]. As a form of economic connectedness, being connected to successful Machine learning role models of an underrepresented background bridges the cultural gaps that increasingly discourage the advancement of students into early career and beyond. In absence of role models, underrepresented students have a harder time imagining themselves working in careers like Machine Learning, creating a cyclical challenge for programs trying to discover talent in the first place, much less cultivate it.

To improve the lifecycle of talent in the figure above, we can (1) retain more representation (of underrepresented researchers) which enable self-conceptual career pathways for prospective students, and (2) continue to identify and fill the cultural/skill gaps, building bridges from both sides of the disparity; in hopes that this, in turn, may create a curbside effect [cite] for collaboration efficacy, providing benefits from inspiration to retention, reducing attrition, and potentially benefiting those along all sides of the divides.

3 Future Work

Identify and study deeper frameworks for ML research impact and collaboration efficacy. Practices like Model Cards and Data Sheets are fairly recent standardizations of preferred ML practices. These types of standards for every stage of an ML researcher’s career, as well as for every type of collaboration. “Standards all the way down”

Rather than imposing culture, finding cultural similarities as we work towards better representation as a means of broadening research collaboration. Colonialization is not support (lol ok, we should reword this!)

Social Science studies like the Difference Education work throughout all stages in scientific careers. Difference Education showed that acknowledging adversity enhances the scholastic performance of racial and socio-economic minorities. As scientists, we know that diversity is paramount to discovery, but how much is it an echo chamber of tradition versus a refined and maturing practice. And as advocates for diversity (whether through our studies or datasets), we may want to regularly evaluate the barriers created through cultural homogeneity of our leading scientific communities. Burden of reconciliation should rest on the in-crowd… [CITE: Is Your Toxicity My Toxicity? Exploring the Impact of Rater Identity on Toxicity Annotation - https://arxiv.org/abs/2205.00501]

4 Conclusion

Once students arrive at their graduate programs, they are often expected to have “made it” [CITE SOMETHING]. Standardization of educational degrees poses a convenience when measuring impact through student retention versus the impact of later careers. We can measure a program’s success by how many of them go on to earn PhDs [Mcnair PhDs]. Post-education, the success measures are less pronounced, which widens the cultural divisions. Filling culture gaps will enable better skill acquisition and a more equitable collaboration ecosystem for ML Making implicit stuff explicit… Standards enable us to fill culture gaps

•

You should directly generate PDF files using pdflatex.
•

You can check which fonts a PDF files uses. In Acrobat Reader, select the menu Files $>$ Document Properties $>$ Fonts and select Show All Fonts. You can also use the program pdffonts which comes with xpdf and is available out-of-the-box on most Linux machines.
•

The IEEE has recommendations for generating PDF files whose fonts are also acceptable for NeurIPS. Please see http://www.emfield.org/icuwb2010/downloads/IEEE-PDF-SpecV32.pdf
•

xfig "patterned" shapes are implemented with bitmap fonts. Use "solid" shapes instead.
•
The \bbold package almost always uses bitmap fonts. You should use the equivalent AMS Fonts:
```
   \usepackage{amsfonts}
```
followed by, e.g., \mathbb{R}, \mathbb{N}, or \mathbb{C} for $\mathbb{R}$ , $\mathbb{N}$ or $\mathbb{C}$ . You can also use the following workaround for reals, natural and complex:
```
   \newcommand{\RR}{I\!\!R} %real numbers
   \newcommand{\Nat}{I\!\!N} %natural numbers
   \newcommand{\CC}{I\!\!\!\!C} %complex numbers
```
Note that amsfonts is automatically loaded by the amssymb package.

Acknowledgments and Disclosure of Funding

Use unnumbered first level headings for the acknowledgments. All acknowledgments go at the end of the paper before the list of references. Moreover, you are required to declare funding (financial activities supporting the submitted work) and competing interests (related financial activities outside the submitted work). More information about this disclosure can be found at: https://neurips.cc/Conferences/2022/PaperInformation/FundingDisclosure.

Do not include this section in the anonymized submission, only in the final paper. You can use the ack environment provided in the style file to autmoatically hide this section in the anonymized submission.

Appendix A Appendix

Optionally include extra information (complete proofs, additional experiments and plots) in the appendix. This section will often be part of the supplemental material.